6 . 1 Introduction
From the data and discussion of both the field trial and the controlled environment study it is quite clear that sainfoin plants display a high level of heterogeneity . The effects of this will be dependant on the obj ectives of a given experiment . It may present problems in that i t reduces the sensitivity o f more conventional field trials , or it may b e mos t useful in providing flexibility t o a plant breeder .
The first point is often alluded to in reviews and comparisons of work ( e . g . Fortune and Withers , 1 980 ) where there is a lack of consistency of trends between experiments . The work of Rumball ( 1 982 ) , on plant introduction trials provides a useful example of using patterns of variation to form certain general groups of l ines from a germplasm collection .
In this study this approach is taken further with the use of multivariate statistical techniques to examine variability in a more orderly and less subj ective manner. There is nothing new in the theoretical and conceptual basis to multivariate analysis but with the increasing availability of computer services the facilities now exist to undertake the large scale numerical manipulations that are often involved .
means clustering. This permitted the examination of the relationships between the measured. variables and also to see whether there is some basis for grouping the individuals into definable lines . The overall objec tive of this numerical approach was to examine the variability , simpl ify the complex array of data collected , and to generate hypotheses on how to make effective use of the reduced data.
6 . 2
The detailed outline of the plant propagation p rocedure was presented in Section 5 . 2 . The plant material that provided the data for this part of the programme was obtained from the plants involved in the redistribution experiment . When the plants had al l been treated with
14 c
at about floral maturity , all top material was removed after 24 hours and dissected . Thus for each of the two cultivars , Fakir and Melrose , there was an information , matrix of 5 6 individuals by 8 variables . The variables were flower number , floral bud number , total dry weight of flowers and floral buds , number of stems , total and individual stem weight , total l eaf weight and specific leaf area.6 . 2 . 1 Data
The data were examined using factor analysis and K-means clustering .
The computer programmes used were part of the BMDP statistical package ( Dixon et al . , 1 98 1 ) on a PRIME 750 computer system.
Factor analysis involves ext rac t i ng factors from the correlation matrix with the aim of reducing the number of variables to a more
manageable number of factors without any loss of information . The underlying assumption is that a smaller number of factors exist that can reproduce exactly the correlations in a larger set of variables . There are a number of methods of extracting factors and these will vary with the computer programme used and personal preference, rather than rigid
statist ical demands . In this experiment a principal components procedure with subsequent varfmax rotation was used . The aim of the rotation was to
achieve a simple structure by maximising the important factor loadings while minimising the others . Ultimately for each case the set of variables can be replaced by a more limited set of factor scores and these examined
for deviations from the mean. The factor scores are derived from the values of each of the variables multiplied by the standard scores for each variabl e of the individual cases. This is in effect the same as using the factor structure as an equation similar to multiple regression to arrive at simple descriptors for each case.
The K-means clustering procedure is an iterative process of allocating various cases
(
individual plants)
to a number of groups based on a criterium of minimising within groups sums of squares . The separation of cases is based on a multivariate d istance measure , in this case theEuclidian distance. The exact number of clusters can be d efined or based
on a descriptive F-test that compares the between-cluster mean square to the within-cluster mean square . All data were initially standardised to z scores
(
standardised normal deviates)
to reduce all attributes to a scale of comparable range , so that the recording scales do not automatically permit the dominance of certain variables .6 . 3 Results
Of the eight principal components shown in Table 6 . 1 , only the first two are maintained for subsequent analysis as these account for about 76% o f the overall varianc e . In this case the first factor explained a little more than 50% of the variance while the second factor add s a further 25% . The exact number of factors used is the result of fairly empirical techniques and in many computer programmes a factor with an eigenvalue
(
v ariance explained)
of less than one is discarded . Low values usually represent either error variance or influences which affect only one or few of the variables . While not shown here, the extraction of a t hird factor separates out SLA which is relatively poorly defined by the first two f actors . Table�
Factor 1 2 3 4 5 6 7 8Complete factor structure for Fakir and Melrose after principal component extraction .
FAKIR MELROSE
Variance % of Variance % of
Explained Total Explained Total
4 . 33 5 4 . 1 4 . 29 5 3 . 6 1 . 76 7 6 . 1 1 . 8 1 76 . 2 0 . 95 8 8 . 0 0 . 79 8 6 . 1 0 . 57 95 . 2 0 . 66 94 . 3 0 . 23 9 8 . 1 0 . 24 97 . 2 0 . 08 9 9 . 1 0 . 1 4 99 . 0 0 . 04 9 9 . 6 0 . 04 99 . 6 0 . 03 1 00. 0 0 . 05 1 00 . 0
Table 6 . 2 illustrates both the factor structure after the initial extraction and after the varimax rotation procedure . It is immediately apparent that the rotation maximised certain of the variables and these have b een underlined in Table 6 . 2 to provide emphasis . On the basis of this , Factor 1 might be regarded as a later maturing, vegetative descriptor , while Factor 2 has a strong loading for flower number and weight . Any reference to maturity was based on the comparative numbers of flowers and buds , rather than an index or scale based on flowering dates . The variable weight/stem seems to be associated with both factors . Both cultivars d isplayed a marked similarity with their overall structure .
While the rotated result made interpretation a little easier , the same result could be derived from the unrotated factors . If they are examined , it can be seen that the first factor might be regarded as a general descriptor of plant size and structure , while the second was a bipolar factor with
o
pposite loadings on the " reproductive" and " vegetative" attributes . Thi s emphasises the fact that rotation was not altering the relationships but was a rescaling procedure .Table 6 . 2
Loadings of the variables on the first 2 factors after both direct extraction and varimax rotation .
FAKIR
2
Variable Direct Rotated Direct Rotated
Flower No . 0 . 6 2 o . 1 8 -0 . 70
.9..:.2..Q.
Bud 0 . 9 0 0 . 78 0 . 02 0 . 45 Flower wt 0 . 74 0 . 33 -0 . 58 0 . 8 8 Stem 0 . 70 0 . 84 0 . 45 -0 . 02 Stem wt 0 . 90�
0 . 36 0 . 1 6 Leaf wt 0 . 84�
0 . 50 0 . 03 Wt/Stem 0 . 70 0 . 57 -0 . 04 0 . 40 SLA -0 . 3 1 0 . 07 0 . 6 4 -0 . 7 1 MELROSEVariable Direct Rotated Direct Rotated
Flower No . 0 . 49 -0 . 06 0 . 80
M!!.
Bud 0 . 9 0 0 . 77 -0 . 08 0 . 45 Flower wt 0 . 67 0 . 1 4 0 . 70�
Stem 0 . 68 0 . 86 -0 . 52 -0 . 03 Stem wt 0 . 9 1M.2.
0 . 25 0 . 32 Leaf wt 0 . 7 7�
-0 . 54 o . oo Wt/Stem 0 . 7 6 0 . 54 0 . 1 3 0 . 5 4 SLA -0 . 60 -0 . 39 -0 . 1 7 -0 . 48Table 6 . 3 provides a g eneral summary of the five clusters formed for Fakir . Five clusters were chosen in each case to provide a balance between descriptive power and an unnecessary number of smal l clusters . The
following notes provide a brief description of the clusters . Table
h..3_
Fakir c lusters and character means ( weights = g/plant ) .
1 2 3 4 5 variable grand mean Flower No . 1 4 . 6 5 . 4 6 . 2 3 . 5 0 . 3 4 . 9 Bud No . 46 . 3 40. 2 70. 6 30 . 3 1 9 . 3 3 4 . 3 Flower wt 5 . 9 2 . 5 4 . 1 1 . 7 0 . 4 2 . 3 Stem No . 9 . 5 1 3 . 8 1 5 . 0 7 . 3 7 . 7 9 . 1 Stem wt 1 1 . 6 1 2 . 5 24 . 0 8 . 4 5 . 7 9 . 9 Wt/Stem 1 . 3 0 . 9 1 . 6 1 . 2 0 . 7 1 • 1 Leaf wt 9 . 4 1 2 . 2 1 7 . 9 7 . 8 6 . 7 9 . 0 SLA 228 282 249 2 27 287 252 Top mass 26 . 9 27 . 2 46 . 0 1 7 . 9 1 2 . 8 2 1 . 2 Cases 1 0 5 5 1 9 1 7 56
Cluster 1 : a group of earlier maturing plants that have a large number of well d eveloped flowers .
Cluster 2 : these plants tend to have a greater than average number of small stems .
Cluster 3 : these plants are later maturing than those in Cluster 1
but still show reproductive prolificacy as well as strong vegetative development .
Cluster 4 : } } Cluster 5 : }
Table 6 . 4
both these clusters are similar and the plants tend to be at , or below the grand mean for most attributes . Cluster 5 does , however, incorporate plants with high SLA ' s and few other distinguishing features except for a constantly poor value for the other variables .
Melrose clusters and character means
(
weights = g/plant)
.1 2 3 4 5 variabl e grand mean Flower No . 3 1 . 0 0 4 . 6 1 . 3 2 . 6 3 . 1 Bud No . 64·. o 1 1 . 6 6 7 . 0 20 . 7 48 . 2 37 . 8 Flower wt 1 1 . 5 0 . 2 4 . 0 1 . 0 2 . 1 2 . 1 Stem No. 1 0 . 0 6 . 8 1 7 . 2 7 . 2 1 3 . 8 1 1 • 1 Stem wt 1 1 . 2 1 . 6 2 1 . 5 3 . 9 9 . 4 8 . 0 Wt/Stem 1 • 1 0 . 3 1 . 3 0 . 6 0 . 7 0 . 7 Leaf wt 9 . 6 5 . 8 1 9 . 3 7 . 5 1 2 . 8 1 0 . 8 SLA 1 98 330 205 2 4 1 234 2 4 1 Top mass 3 2 . 3 7 . 6 44 . 8 1 2 . 4 24 . 3 20 . 9 Cases 2 5 5 1 9 25 5 6
Table 6 . 4 provides a general summary of t h e five clusters formed for Melrose. The following notes provide a b rief d escription of the clusters .
Cluster 1 : only two plants well separated from the other clusters by their degree of floral d evelopment and maturity .
Cluste r 2 : these plants are below average for most attributes but
Cluster 4 : } } Cluster 5 : }
except SLA. All plants have a large number of well
developed stems which p rovide sites for b oth leaf and floral development , and this is reflected in the
respective values for leaf weight and bud number .
two more general clusters where attributes are distributed either slightly above ( Cluster 4 ) , or below
( Cluster 5 ) , the grand mean values .
Total top mass of the plants in each cluster is shown in Tables 6 . 3 and 6 . 4 , but was not used as a variabl e in the clustering procedure . Thi s v alue was p rovided a s an indicator o f harvestable production from various clusters .
As the two cultivars showed similar patterns with the type of factors and clusters formed it was important to consider any differences that may have occured between them. Table 6 . 5 shows that Melrose produced more stems than Fakir and more leaf mass at or near maturity . The smaller average stem s ize on the Melrose plants is balanced by the greater number with the resul t being that both cultivars have a similar stem mas s . I t should b e noted that Melrose was harvested about 7 days later than Fakir and this may have contributed to its advantage in leaf production .
Table
U
Comparisons between Melrose and Fakir where differences occur. ( p<0 . 05 ) Fakir C±S . E. ) Melrose C±S . E . )
Stem No . 9 . 08 ( 0 . 48 ) 1 1 . 07 ( 0 . 64 )
Wt/Stem 1 . 07 ( 0 . 05 ) 0 . 69 ( 0 . 05 )
Leaf Weight 9 . 04 ( 0 . 53 ) 1 0 . 85 ( 0 . 6 1 )
Table 6 . 6
Factor scores for individual cases associated with some of the c lusters . Factor Score
Cultivar Cluster Case 1 2
Melrose 1 6 -0 . 62 5 . 1 2 2 5 -0 . 57 2 . 27 3 1 6 2 . 06 0 . 05 40 2 . 97 -0 . 53 48 2 . 30 0 . 54 Fakir 1 5 0 . 07 2 . 27 4 1 -0 . 73 1 . 9 8 47 -0 . 0 1 1 . 7 6 3 22 2 . 06 0 . 79 33 3 . 67 -1 . 3 4 3 5 1 . 86 -0 . 0 2
Table 6 . 6 provides some indication o f the comparative outcome of the two techniques . It can be seen that the factor scores do give an indication as to the cluster arrangement that was revealed by cluster
analysis . Cluster 1 in both cultivars are plants with well developed flowers and so they also rate high factor scores for Factor 2 . Conversely the plants indicated in Cluster 3 are late maturing and have above average vegetative development and so they score highly for Factor 1 .
6 . 4 Discussion
The object of this examination of data associated with two sainfoin cultivars was not to make definitive statements about individual
performance , but rather to examine patterns of variation that exist , and how these may be usefully exploited . While it may be argued that many of the patterns observed , for the way in which both variables and cases tend to group , could have been formed on an intuitive basis from examination of means , variances , and correlations , the multivariate techniques permit a formalisation of this approach . It should however , b e emphasised that the overall success ' of these and other multivariate techniques are very dependant on a good understanding of the strengths and weaknesses of the techniques . In this way these methods are the same as the more familiar univariate statistic s ( Pearce , 1 9 69 ) .
Given these introductory perspectives , the first point of importance is that for the factor analysis there has been a data reduction from 448 ( 56x8 ) elements to 1 1 2 ( 56x2 ) elements , and for the K-means clustering from the same start to 40 element s ( 5x8 ) and the 8 grand mean values . While data reduction and summary are not confined to multivariate techniques , these procedures do permit a rapid and structured examination of the dat a . In this case the examination will be confined to some general aspects o f the results and one o r two more specific points which might be regarded as
hypothesis generating.
Of particular interest is the extent of variation that occurs within both of the named cult ivars . While it is accepted that an open poll inated population will be variable
(
Simmonds , 1 979 ) , cultivar development usually involves some narrowing of the d istribution . This is especially so for characters such as flowering time which tend to be quite amenable to selection i . e . character heritability is high . However, in this study , b oth techniques use d egree of floral d evelopment and maturity as key criteria on which t o d ifferentiate plants . Fakir and Melrose produc e d istinctive clusters(
Cluster 3 ) , of later maturing plants that have s trong vegetative development in terms of leaf and stem , a s well as the potential for reproductive yield as indicated by the high number of developing floral buds . Cluster 1 for both cult ivars are plants that show strong early floral development , and only average vegetative yield . Factor analysis t ends to result in similar grouping patterns as the factors retained weight for either late maturing, vegetative attributes or early floral development . This range of flowering time also means that under field conditions the choice of a single occasion when all plants are at or near some optimum point o f development is highly improbable.It is the Cluster 3 plants that would seem to offer some avenue for gain as they appear to have the combination of good vegetative yield , total harvestable yield and t he potential for seed p roduction , if factors other than flower numbers are not limiting . The first features are the essential basis of a forage plant while the latter would ensure ready dissemination of any improved material .
While this study is not extensive enough to confirm or deny the hypotheses developed elsewhere ( e . g . Pearce et 1 969 ; Straley et al . ,
1 972 ) , that modification of SLA may effect photosynthetic efficiency and t herefore production , it is of interest to examine Cluster 5 ( Faki r ) and Cluster 3 ( Melros e ) . Both these clusters have plants with well above average SLA ' s that show poor performance for all other characters , especially leaf mass . Therefore some caution may b e necessary with the suggestion of Sheehy and Popple ( 1 9 8 1 ) that leaf area index of sainfoin might be improved by selecting for higher SLA .
Plants with poor flower and bud development , as for example those in Melrose Cluster 2 , t end t o b e well below average in terms o f stem numbe r . While this may reflect a general l ack o f vigour , it also illustrates the importance of stem numbers in providing the necessary sites for both flower and leaf d evelopment .
.
It is obvious that further ideas can b e generated from the data , but to attempt excessiv e analysis or conclusions from the limited population would to some extent be an abuse of the techniques . Therefore , the next step would be to examine other populations to see if their variation conforms with the general patterns observed here . Also select groups of plants that have been indicated as having certain key attributes , such as a high number of s tems and a l ater than average flowering time , might be studied in greater detail . If the two cultivars examined in this study are regarded as a representative subset of a larger population , probabilities of finding certain combinations of attributes could also be derived giving
some indication of how possible gains might be without resorting to
would also require heritability estimates of the various desirable traits
(
Varga ,1 968 ;
Simmo.nds , 1 979 ) .In conc lusion , if there is reasonable access to computing facilities the multivariate techniques offer the opportunity to readily examine the structure of a population and the interactions amongst measured variables . This can reveal certain patterns that can be directly useful in selecting plants and also provide the basis for further hypothesis generation . For the relatively limited sainfoin data base used in this study , groupings of plants were formed that were consistent with observations from the field e . g . later flowering time and higher yield . However, the magnitude of the differences that occured between the groupings of plants was revealing , and offers considerable scope for population improvement if the differences are shown to be consistent on a larger sample .
CHAPTER
�
GENERAL DISCUSSIONIn the field and controlled environment s tudies reported in this thesis sainfoin was examined to determine factors that were likely to influence the pattern and extent of growth when subject defoliation . Different cultivars were used to provide variation in flowering time and expected regrowth . ability. Without attempting to preempt unbiased evaluation of sainfoin in light of this work , it is important that due recognition be given to the fact that the germplasm base for the experiments was quite narrow, and that the local environmental conditions were favourable for lucerne production .
The s tudy was undertaken on the b asis that sainfoin was a highly desirable forage legume in terms of its feeding value to ruminants . However, a forage legume must also possess a number of other attributes that are prerequisites for successful commercial adoption . The plant must b e easy to establish , must readily form an effective symbiotic
relationship , b e highly productive under a range of management treatments , be persistent , have a good pest and disease resistance , and ultimately produce enough seed to enable cheap dissemination to users ( Rogers , 1 975 ; Humphreys , 1 98 1 ; Vickery , 1 98 1 ) . By considering these various aspects with reference to the results and observations reported in this thesis , it should become clear where the relative strengths and weaknesses of sainfoin appear to lie .
Sainfoin proved to have the ability to germinate rapidly and to
have early growth rates that were comparable with lucerne. Apart from some
sainfoin was a new crop to the area , the procedures for establishing sainfoin were conventional and did not require special equipment . The ready establishment of sainfoin has also been noted in other experiments ( Smoliak et 1 972 ; Cooper , 1 977 ; Smoliak and Hanna , 1 977 ) , and is often attributed to the large seed size of sainfoin and its ability to grow rapidly at relatively low temperature s . The drawback to the large seed size is that the recommended sowing rates to achieve a similar starting population to other legumes is high ( Spedding and Diekmahns , 1 972 ; Fortune and Withers , 1 980 ) . Currently this makes sainfoin expensive to establish